tribev2 0.0.2

TRIBE v2 — multimodal fMRI brain encoding model inference in Rust
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220

/// Fused CubeCL GPU kernels for the TRIBE v2 encoder.
///
/// **ScaleNorm** — single-pass L2-normalise + scale using `plane_sum`
///   (Metal SIMD-group all-reduce, plane_dim = 64).
///   Replaces: pow → sum_dim → clamp → pow → mul×2 (≥3 dispatches × 17 calls).
///
/// **RoPE rotation** — one kernel dispatch per Q or K application.
///   Replaces: slice×3 + mul×4 + sub + add + cat (≥5 dispatches + 4 allocs × 16 applies).

use cubecl::prelude::*;
use cubecl::wgpu::WgpuRuntime;
use cubecl::server::Handle;
use burn::backend::wgpu::CubeTensor;
use burn_backend::{DType, Shape};
use burn_cubecl::kernel::into_contiguous;

// ── tensor helpers ────────────────────────────────────────────────────────

fn elem_size(dtype: DType) -> usize {
    match dtype {
        DType::F64 | DType::I64 | DType::U64 => 8,
        DType::F32 | DType::Flex32 | DType::I32 | DType::U32 => 4,
        DType::F16 | DType::BF16  | DType::I16  | DType::U16 => 2,
        DType::I8  | DType::U8   | DType::Bool => 1,
        DType::QFloat(_) => 4,
    }
}

fn contiguous_strides(dims: &[usize]) -> Vec<usize> {
    let n = dims.len();
    let mut s = vec![1usize; n];
    for i in (0..n.saturating_sub(1)).rev() {
        s[i] = s[i + 1] * dims[i + 1];
    }
    s
}

fn empty_cube(src: &CubeTensor<WgpuRuntime>, shape: Shape) -> CubeTensor<WgpuRuntime> {
    let n_bytes  = shape.dims.iter().product::<usize>() * elem_size(src.dtype);
    let handle: Handle = src.client.empty(n_bytes);
    CubeTensor {
        client:  src.client.clone(),
        device:  src.device.clone(),
        handle,
        strides: contiguous_strides(&shape.dims),
        shape,
        dtype:   src.dtype,
        qparams: None,
    }
}

// ── ScaleNorm ─────────────────────────────────────────────────────────────
//
// Grid:   CubeDim { x: 64, y: 1, z: 1 },  CubeCount::Static(rows, 1, 1)
// CUBE_POS_X = which row (u32),  UNIT_POS_X = lane within SIMD-group (u32)
//
// Scalar kernel args must be Rust primitive types (u32/f32/usize), NOT `F: Float`
// (F is a CubeCL IR type — it doesn't exist at the Rust host level).

#[cube(launch)]
fn scalenorm_kernel<F: Float>(
    x:     &Tensor<F>,
    g:     &Tensor<F>,           // [1]
    out:   &mut Tensor<F>,
    d:     u32,                  // last dim size (e.g. 1152)
    scale: f32,                  // sqrt(D), cast to F inside
) {
    let row  = CUBE_POS_X;       // u32
    let lane = UNIT_POS_X;       // u32  (0..PLANE_DIM)

    let n_per_lane = d / PLANE_DIM;
    let base = row * d + lane;

    let mut sq = F::new(0.0);
    for i in 0u32..n_per_lane {
        let v = x[(base + i * PLANE_DIM) as usize];
        sq += v * v;
    }

    let row_sq = plane_sum(sq);
    let factor = g[0usize]
        * F::cast_from(scale)
        * F::powf(F::max(row_sq, F::new(1e-12)), F::new(-0.5));

    for i in 0u32..n_per_lane {
        let idx  = (base + i * PLANE_DIM) as usize;
        out[idx] = x[idx] * factor;
    }
}

pub fn launch_scalenorm(
    x:     CubeTensor<WgpuRuntime>,
    g:     CubeTensor<WgpuRuntime>,
    scale: f32,
) -> CubeTensor<WgpuRuntime> {
    let x   = into_contiguous(x);
    let g   = into_contiguous(g);
    let out = empty_cube(&x, x.shape.clone());

    let d    = *x.shape.dims.last().unwrap() as u32;
    let rows = x.shape.dims[..x.shape.num_dims() - 1].iter().product::<usize>() as u32;

    // One plane (64 threads on Metal) per row
    let cube_dim   = CubeDim { x: 64, y: 1, z: 1 };
    let cube_count = CubeCount::Static(rows, 1, 1);

    match x.dtype {
        DType::F32 | DType::Flex32 => unsafe {
            scalenorm_kernel::launch::<f32, WgpuRuntime>(
                &x.client, cube_count, cube_dim,
                TensorArg::from_raw_parts::<f32>(&x.handle,   &x.strides,   &x.shape.dims, 1),
                TensorArg::from_raw_parts::<f32>(&g.handle,   &g.strides,   &g.shape.dims, 1),
                TensorArg::from_raw_parts::<f32>(&out.handle, &out.strides, &out.shape.dims, 1),
                ScalarArg::new(d),
                ScalarArg::new(scale),
            ).expect("scalenorm f32 launch");
        },
        DType::F16 => unsafe {
            scalenorm_kernel::launch::<half::f16, WgpuRuntime>(
                &x.client, cube_count, cube_dim,
                TensorArg::from_raw_parts::<half::f16>(&x.handle,   &x.strides,   &x.shape.dims, 1),
                TensorArg::from_raw_parts::<half::f16>(&g.handle,   &g.strides,   &g.shape.dims, 1),
                TensorArg::from_raw_parts::<half::f16>(&out.handle, &out.strides, &out.shape.dims, 1),
                ScalarArg::new(d),
                ScalarArg::new(scale),
            ).expect("scalenorm f16 launch");
        },
        dt => panic!("scalenorm: unsupported dtype {dt:?}"),
    }
    out
}

// ── RoPE rotation ─────────────────────────────────────────────────────────
//
// Grid: CubeDim { x: 256, y: 1, z: 1 },  CubeCount::Static(ceil(total/256), 1, 1)
// ABSOLUTE_POS (usize) = global thread index across the whole B×H×N×D tensor.
//
// For usize scalar args (ABSOLUTE_POS is usize, so arithmetic with total/d/half
// uses usize for consistency).

#[cube(launch)]
fn rope_kernel<F: Float>(
    x:     &Tensor<F>,
    cos:   &Tensor<F>,           // [N, half]
    sin:   &Tensor<F>,           // [N, half]
    out:   &mut Tensor<F>,
    d:     usize,                // head dim (144)
    half:  usize,                // D/2      (72)
    n_seq: usize,                // seq len  (100)
    total: usize,                // B×H×N×D
) {
    let pos = ABSOLUTE_POS;
    if pos >= total { terminate!(); }

    let d_idx = pos % d;
    let n_idx = (pos / d) % n_seq;

    let result: F = if d_idx < half {
        let cs = n_idx * half + d_idx;
        x[pos] * cos[cs] - x[pos + half] * sin[cs]
    } else {
        let cs = n_idx * half + d_idx - half;
        x[pos - half] * sin[cs] + x[pos] * cos[cs]
    };
    out[pos] = result;
}

pub fn launch_rope(
    x:   CubeTensor<WgpuRuntime>,
    cos: CubeTensor<WgpuRuntime>,
    sin: CubeTensor<WgpuRuntime>,
) -> CubeTensor<WgpuRuntime> {
    let x   = into_contiguous(x);
    let cos = into_contiguous(cos);
    let sin = into_contiguous(sin);
    let out = empty_cube(&x, x.shape.clone());

    let dims  = &x.shape.dims;
    let rank  = dims.len();
    let d     = dims[rank - 1];
    let half  = d / 2;
    let n_seq = dims[rank - 2];
    let total = dims.iter().product::<usize>();

    let cube_dim_x: u32 = 256;
    let cube_count_x    = ((total as u32) + cube_dim_x - 1) / cube_dim_x;
    let cube_dim        = CubeDim { x: cube_dim_x, y: 1, z: 1 };
    let cube_count      = CubeCount::Static(cube_count_x, 1, 1);

    match x.dtype {
        DType::F32 | DType::Flex32 => unsafe {
            rope_kernel::launch::<f32, WgpuRuntime>(
                &x.client, cube_count, cube_dim,
                TensorArg::from_raw_parts::<f32>(&x.handle,   &x.strides,   &x.shape.dims, 1),
                TensorArg::from_raw_parts::<f32>(&cos.handle, &cos.strides, &cos.shape.dims, 1),
                TensorArg::from_raw_parts::<f32>(&sin.handle, &sin.strides, &sin.shape.dims, 1),
                TensorArg::from_raw_parts::<f32>(&out.handle, &out.strides, &out.shape.dims, 1),
                ScalarArg::new(d),
                ScalarArg::new(half),
                ScalarArg::new(n_seq),
                ScalarArg::new(total),
            ).expect("rope f32 launch");
        },
        DType::F16 => unsafe {
            rope_kernel::launch::<half::f16, WgpuRuntime>(
                &x.client, cube_count, cube_dim,
                TensorArg::from_raw_parts::<half::f16>(&x.handle,   &x.strides,   &x.shape.dims, 1),
                TensorArg::from_raw_parts::<half::f16>(&cos.handle, &cos.strides, &cos.shape.dims, 1),
                TensorArg::from_raw_parts::<half::f16>(&sin.handle, &sin.strides, &sin.shape.dims, 1),
                TensorArg::from_raw_parts::<half::f16>(&out.handle, &out.strides, &out.shape.dims, 1),
                ScalarArg::new(d),
                ScalarArg::new(half),
                ScalarArg::new(n_seq),
                ScalarArg::new(total),
            ).expect("rope f16 launch");
        },
        dt => panic!("rope: unsupported dtype {dt:?}"),
    }
    out
}